Image registration is one of fundamental researching areas in computer vision, which is used to determine the mapping between two images both spatially and with respect to intensity, for purpose of integrating the information from both images. There are two groups of registration methods with respect to the type of spatial transformation: global models and local method. Local method can deal with complex variations and distortions between images, and match them more exactly than global methods. The elastic matching model is one of frequently studied local methods. In this paper we proposed a registration approach based on the simplified linear elastic model. First the image are globally aligned with a principle axis method, then a rectangular elastic net is applied on the images for local matching. Every nodes on the net received external force from its corresponding point, and internal force from its four neighbors. The elastic net will deform and finally reach a balanced state. This method takes a multi-resolution searching strategy for the corresponding points in order to archive better matching. We also proposed a mapping function using on cubic spline interpolation, which is very fast and considerable precise. We also take some work on elastic matching based on non-linear diffusion image scale-space, by introducing a scale-related diffusion parameter. Many hierarchical strategies are introduced into image registration to avoid local minima trap, methods base on multi-scale images is one of them, which decrease image details gradually. In recent years, non-linear scale-space gained more and more attention, for its ability of preserving import features. We make an introduction to non-linear scale-space, and apply it into the elastic registration.
修改评论